We are all starting to have AI in our lives every day, from Amazon recommendations to improved picture-taking ability on our mobile phones. Yet AI can be a game-changer for the process industry as well.

Why AI for Processing Plants is Different From Consumer AI

The most obvious example of AI in everyday life is e-commerce Web sites that recommend items based on “Customers who viewed this item also viewed…” This is done by analyzing the browsing and purchase processes of millions of visitors in order to make recommendations. However what Amazon and other retailers do not take into consideration is that there are other factors that will influence each purchase, such as a spouse or friends, or whether the buyer has had lunch or not yet.

Unlike online purchases, in systems like process plants it is possible to take into consideration most influencing factors such as changes in valves, pressure vessels, catalysts and reactors. That is, almost every aspect of the plant is known, every process, every piece of equipment, every raw material, temperature, pressure, and so on, and data is continuously tracked and stored in the historian. When using Precognize, our unique methodology starts by modeling the entire plant, and then combining that with all of the plant’s tracked data. This is extremely powerful when you realize that every process plant can already have a vast piece of the AI puzzle in its hands, something that even the great AI leader Amazon does not have.


That said, Amazon has an easier job in that they have access to the purchase behavior of millions of people from which to build their models. In the case of a process plant, a wide range of data exists, but the number of problems – which is the data element that you want to track – is (typically, anyway) quite low. So you cannot build your AI algorithm by tracking past problems and using these to predict the next problems; this simply will not work.

So how we do it at Precognize is by analyzing the data of the entire plant using machine learning to understand what normal operation looks like. Once we have a strong baseline of the normal behavior, it’s much easier to determine when something deviates from that norm, i.e. an anomaly in the data.

However, it is not enough to identify that something in the plant is behaving in a way outside the norm; we need to be able to determine if the anomaly is also meaningful. How to do this is another process altogether, which involves combining AI with the knowledge of plant engineers.

People are Still Needed, Even with AI

Through our extensive work in the process industry, we have determined that the most efficient way to determine the meaningfulness of a particular anomaly is to build an ontology, a sort of a dictionary or map that describes the entire plant, all of its equipment and processes. It is comprised of the plant engineers’ knowledge about the plant, such as its structure and processes and then be enriched by linking between the things that influence each other, and then used by the AI algorithms to determine the level of meaning of each alert.

This precise mathematical model of the plant can be accomplished within around two weeks, even in very large plants. This method of augmenting the AI makes it much more valuable to the plant operators since the AI can “understand” what is linked to what, what influences what. And the AI algorithms can deal with the vast quantities of data the plant is generating, which human beings could not.

This model of the plant gets turned into a mathematical graph and we can apply AI techniques to it. This enables a minimum of alerts, without overloading operators with noise.

Asking the Right Questions from the Start

To get the most out of AI technology, you need to start with a non-technical question: what are you trying to achieve in your plant? What are the KPIs, and what data and knowledge actually allows you to understand the root causes of problems, and to fulfill the designated KPIs?

We recommend to start any AI implementation with an impact analysis, even before utilizing a single unit of code. This early phase usually takes as little as one week, but can make the difference between your AI project’s success or failure.

Once the expectations are set in this early phase, you can begin the modeling workshop, which enables engineers to share their knowledge to build the dictionary or map of the plant that can be injected into the AI. Then it is just a matter of a few days to review early results and fine-tune the model structure or processes, until the system can go live, and true value can be generated.

AI: Not Just for Technology’s Sake

While there’s been hype around AI in recent years, it’s important to understand that it cannot be a quick fix to any one problem. As such, it shouldn’t be applied to a single piece of equipment, but to the system as a whole. This is our approach at Precognize.

It will also not, as has been rumored, eliminate human involvement, at least not in the coming decade or more. Employees are still integral, especially in the processing industry, which needs their knowledge both to apply to the AI itself, and also to remediate issues as they are flagged by the AI system.

The real power of AI, we now know, is in its ability to crunch massive amounts of data, and to incorporate knowledge from people on the processing plant floor. We’ve seen collaboration between the technology and human knowledge result in impressive AI algorithms that are able to learn, and relearn, from production patterns that are always in flux. Such a dynamic system is necessary to ensure that technology is working for your plant— to increase your KPIs, your goals, and your overall output—rather than being implemented, as we’ve seen in the past, simply for the sake of technological advancement.

Learn more about how AI can positively impact your process plant in this webinar.